Cross-Species Multitask Learning with Molecular and ADME Descriptors for Liver Microsomal Metabolic Stability

Area of Science:

• * Computational Chemistry
• * Cheminformatics
• * Drug Discovery

Background:

• * Liver microsomal metabolic stability is crucial for predicting *in vivo* drug exposure and is a key filter in lead optimization.
• * Cross-species prediction of metabolic stability is hindered by varied metabolic pathways and limited model interpretability.
• * Existing models often struggle to generalize across different species (human, rat, mouse).

Purpose of the Study:

• * To develop a cross-species multitask learning framework for predicting liver microsomal metabolic stability.
• * To integrate diverse molecular representations (SMILES fingerprints, molecular graphs) and *in silico* ADME/physicochemical descriptors.
• * To enhance model interpretability by identifying key structural features and substructures influencing stability.

Main Methods:

• * Curated 18,921 PubChem BioAssay measurements for human (HLM), rat (RLM), and mouse (MLM) liver microsomes.
• * Employed a multitask learning framework integrating Morgan, MACCS/RDKit fingerprints, molecular graphs, and ADME descriptors.
• * Utilized stratified 10-fold Bemis-Murcko scaffold cross-validation, ensemble prediction, and species-specific thresholds.

Main Results:

• * Achieved high predictive performance with AUROC values of 0.811 (HLM), 0.806 (RLM), and 0.794 (MLM), and AUPR values of 0.854, 0.860, and 0.862, respectively.
• * The multitask model consistently outperformed conventional machine learning and single-task deep learning baselines.
• * SHapley Additive exPlanations (SHAP) and EdgeSHAPer identified key features (e.g., lipophilicity, CYP interaction) and substructures (e.g., alkenes, amides) influencing stability.

Conclusions:

• * The proposed multi-modal, cross-species framework significantly improves metabolic stability prediction accuracy and interpretability.
• * Integrating diverse molecular encodings with ADME descriptors offers a powerful approach for drug optimization.
• * The framework provides hypothesis-generating insights into structure-metabolism relationships, guiding experimental validation.